save initial state in sampler warmup?

[Bob Carpenter]

On 8/21/13 2:09 PM, Marcus Brubaker wrote:
> Taking a quick look at the code, I don't think the initial state is output during sampling. It would be good to change
> that as well.

I agree.

They're not saved in Stan 1.3 from the command line.
But they are saved in RStan 1.3.

I think it would be nice to have the first values
saved should be the initial ones supplied by the user or
generated randomly.

Michael --- what's the behavior in arg_refactor?

- Bob

[Michael Betancourt]

The behavior is the same as in develop, but any changes to command.hpp in develop will be a pain to merge with the significantly changed version in arg_refactor. It'd be much easier to make the changes directly -- what exactly do people want to see?

> --
> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.
> For more options, visit https://groups.google.com/groups/opt_out.

[Bob Carpenter]

On 8/21/13 3:24 PM, Michael Betancourt wrote:
> The behavior is the same as in develop, but any changes to command.hpp in develop will be a pain to merge with the significantly changed version in arg_refactor. It'd be much easier to make the changes directly -- what exactly do people want to see?

Just what I was thinking.

Marcus is adding saving of initial states in optimization.

We'd like to have user inits saved as the first warmup sample.

We have to explain, too, that we have a procedure to go from
inits to a more reasonable starting point, then from that
starting point we adapt during warmup.

- Bob

[Michael Betancourt]

> Just what I was thinking.
>
> Marcus is adding saving of initial states in optimization.
>
> We'd like to have user inits saved as the first warmup sample.

Both are easy (really, just changing an index around).

> We have to explain, too, that we have a procedure to go from
> inits to a more reasonable starting point, then from that
> starting point we adapt during warmup.

To what "go from inits to a more reasonable starting point" are you referring? Right now the samplers will just start warmup at the user-defined init and with intermediate procedure.

[Bob Carpenter]

I was thinking about the method on NUTS that was called
something like find_reasonable_parameters(). It was called
right in the constructor if I recall correctly.
Maybe that was only finding a reasonable stepsize and
didn't move the parameter values.

It looks like the MCMC refactor got rid of this step.

- Bob

[Matt Hoffman]

Yup, "parameters" there refers to the step size. (Perhaps confusingly.)

Did the refactor eliminate this completely? I think it's kind of
important. Otherwise you might start out with too small a step size
and have a really, really slow first iteration.

Matt

[Michael Betancourt]

Nope, just refactored. Same algorithm.